A forensic science cross-age comparison of children's fingerprint growth system

Abstract

This invention relates to the field of fingerprint recognition technology in forensic science, and discloses a fingerprint growth system for children for cross-age comparison in forensic science. The system includes a fingerprint age normalization generation unit and a multi-physical constraint loss optimization unit. The fingerprint age normalization generation unit adopts a multi-scale encoder-decoder architecture, configured to perform inverse physical modeling of the non-uniform physical deformation of the fingerprint skin as it changes with physiological age on a pre-processed standard fingerprint image of a young child. This fingerprint growth system for children for cross-age comparison in forensic science, through the closed-loop control of the fingerprint age normalization generation unit combined with the multi-physical constraint loss optimization unit, overcomes the limitation of existing explicit geometric transformation correction methods that can only handle globally uniform deformation. The encoder extracts the fingerprint orientation field and spatial deformation features, and the decoder completes the reconstruction of the adult-shaped fingerprint image. The multi-physical constraints adapt to the non-uniform physiological growth law of children's fingerprints, accurately correcting local deformations.

Description

Technical Field

[0001] This invention relates to the field of forensic fingerprint recognition technology, specifically a children's fingerprint growth system for cross-age comparison in forensic science. Background Technology

[0002] The pattern and direction of fingerprint ridges are permanent after embryonic development, and are one of the core biometric features for individual identification in forensic science. However, the persistence of fingerprints is not the same as absolute permanence. Microscopic metric attributes such as ridge spacing and absolute spatial coordinates of detail feature points will gradually change with human growth. This change is more drastic in young children who are in the rapid growth and development stage. Children's fingerprints are also characterized by shallow ridges, small area, and significant deformation of features with age, which makes fingerprint extraction and cross-age fingerprint comparison and identification difficult.

[0003] In forensic science practice, the need for cross-age fingerprint comparison is widespread in scenarios such as the identification of missing children, long-pending cases, and the verification of transnational immigration identities. Chinese law enforcement agencies have accumulated a large number of fingerprint samples from abducted children at a young age during investigations of trafficking cases, which urgently need to be compared with their adult fingerprints. However, traditional automatic fingerprint recognition systems are designed by defaulting to input fingerprints in an adult state. When there is a significant age difference between the queried fingerprint and the registered fingerprint, the matching performance deteriorates significantly, making it difficult to meet the actual needs of case handling. Therefore, how to effectively model and compensate for fingerprint morphological changes caused by growth and development to improve the accuracy of cross-age fingerprint comparison has become a core problem that urgently needs to be solved in the intersection of forensic science and pattern recognition.

[0004] Existing technical solutions for cross-age fingerprint comparison mainly fall into two categories:

[0005] The first category is correction methods based on explicit geometric transformations, which compensate for fingerprint scale differences through global scaling, affine transformations, etc. However, they can only handle global uniform deformations and cannot adapt to local deformations caused by the non-uniform physiological growth of children's fingerprints. Furthermore, simple geometric scaling can easily exacerbate image blurring, leading to an increase in false features, and cannot meet the stringent requirements of forensic science for the authenticity of detailed features.

[0006] The second category is the general deep generative model, which uses generative adversarial networks to achieve image conversion; however, it lacks specific constraints on the physical laws of fingerprint growth, and is prone to generating false detail features during the generation process. At the same time, its optimization objective lacks a direct mathematical connection with the detail point matching logic of the automatic fingerprint recognition system. Under the condition of extremely low false recognition rate that forensic science focuses on, the true matching rate is seriously insufficient and cannot meet the application requirements of criminal investigation scenarios.

[0007] Therefore, a children's fingerprint growth system for cross-age comparison in forensic science is proposed. Summary of the Invention

[0008] (a) Technical problems to be solved

[0009] To address the shortcomings of existing technologies, this invention provides a children's fingerprint growth system for cross-age comparison in forensic science. It possesses advantages such as accurately adapting to the non-uniform physiological growth patterns of children's fingerprints, effectively correcting local deformations in fingerprints across age ranges, strictly suppressing false details during the generation process, significantly improving the true positive rate of fingerprint matching in scenarios with extremely low false recognition rates, and seamlessly adapting to mainstream automatic fingerprint recognition systems in public security and criminal investigation. This invention solves the problems of existing technologies, such as geometric transformation methods only handling globally uniform deformations, easily exacerbating image blurring and increasing false features, and general deep generation models lacking specific constraints on the physical laws of fingerprint growth, easily generating false details, and the optimization target being disconnected from the matching logic of automatic fingerprint recognition systems, thus failing to meet the stringent application requirements of forensic science.

[0010] (II) Technical Solution

[0011] To achieve the above objectives of accurately adapting to the non-uniform physiological growth pattern of children's fingerprints, effectively correcting local deformation of fingerprints across ages, strictly suppressing false detail features during the generation process, significantly improving the true positive rate of fingerprint matching in scenarios with extremely low false recognition rates, and seamlessly adapting to mainstream automatic fingerprint recognition systems for public security and criminal investigation, this invention provides the following technical solution: a children's fingerprint growth system for cross-age comparison in forensic science, comprising a fingerprint age normalization generation unit and a multi-physical constraint loss optimization unit;

[0012] The fingerprint age normalization generation unit adopts a multi-scale encoder-decoder architecture and is configured to perform inverse physical modeling of the non-uniform physical deformation of fingerprint skin with physiological age on the pre-processed standard fingerprint image of young children, and output a normalized fingerprint image corresponding to the adult form.

[0013] The multi-physical constraint loss optimization unit has its input end connected to the output end of the fingerprint age normalization generation unit, and its output end forms a closed-loop hardware connection with the parameter control end of the fingerprint age normalization generation unit. The multi-physical constraint loss optimization unit is configured to construct a composite loss function based on the physical laws of fingerprint physiological growth, including orientation field perception constraints, ridge frequency evolution constraints, and identity preservation constraints. It combines pixel reconstruction loss to generate total optimization parameters, and uses closed-loop feedback to adjust the network parameters of the fingerprint age normalization generation unit, correcting the non-uniform growth deformation of children's fingerprints and suppressing forged minutiae in the generation process.

[0014] Preferably, it also includes a local texture discrimination unit;

[0015] The first input terminal of the local texture discrimination unit is hardware-connected to the output terminal of the fingerprint age normalization generation unit, the second input terminal is connected to a real adult fingerprint standard image, and the output terminal is hardware-connected to the multi-physical constraint loss optimization unit.

[0016] The local texture discrimination unit adopts the PatchGAN architecture, which consists of multiple convolutional layers connected in series. Each convolutional layer is followed by an instance normalization layer and a LeakyReLU activation function. The architecture finally outputs an N×N truth probability matrix.

[0017] The local texture discrimination unit is configured to perform authenticity discrimination on the local ridge structure of the generated normalized fingerprint image and the real adult fingerprint image, apply adversarial penalties to ridge breaks and false textures in the local receptive field, ensure the topological continuity of the local ridges in the generated image, output the local authenticity probability matrix and adversarial loss, and transmit them to the multi-physical constraint loss optimization unit to participate in the generation of the overall optimization parameters. Each element of the probability matrix corresponds to the authenticity probability of a local receptive field of the input fingerprint image.

[0018] Preferably, the orientation field perception constraint of the multi-physics constraint loss optimization unit calculates the orientation field feature mappings of the generated normalized fingerprint image and the real adult fingerprint image respectively through the fingerprint orientation field feature extractor, calculates the mean L1 distance of the two feature mappings, applies gradient penalty to the pixel distortion that deviates from the original fingerprint skeleton flow direction, and locks the topological invariance of the fingerprint macro-pattern.

[0019] Preferably, the ridge frequency evolution constraint of the multi-physical constraint loss optimization unit is obtained by extracting local ridge frequency features, acquiring the local ridge frequency matrices of the generated normalized fingerprint image and the original child fingerprint image respectively, and combining them with the anisotropic dynamic frequency attenuation coefficient matrix prefitted based on the child fingerprint growth law, calculating the mean L1 distance of the two frequency matrices at the corresponding spatial positions. The constraint generation process conforms to the anisotropic attenuation law of the ridge density of the child fingerprint with age.

[0020] Preferably, the identity preservation constraint of the multi-physical constraint loss optimization unit is achieved by extracting high-dimensional identity features of the generated normalized fingerprint image and the real adult fingerprint image through a pre-trained fingerprint identity feature extraction network, calculating the cosine distance between the two high-dimensional identity features, and constraining the identity-related features of the generated image to remain consistent with the real target image.

[0021] Preferably, the total loss function of the multi-physics constraint loss optimization unit is obtained by weighting and summing adversarial loss, pixel reconstruction loss, orientation field perception loss, identity preservation loss, and ridge frequency evolution loss with preset weight coefficients having different values.

[0022] Preferably, it also includes a fingerprint image preprocessing unit;

[0023] The output of the fingerprint image preprocessing unit is hardware-connected to the input of the fingerprint age normalization generation unit. It is configured to sequentially perform histogram equalization, Gaussian filtering, and Laplacian sharpening on the input standard fingerprint image of a child, and output the preprocessed fingerprint image. The histogram equalization is used to expand the global grayscale dynamic range of the fingerprint image, the Gaussian filtering is used to smooth the high-frequency noise of the image, and the Laplacian sharpening is used to enhance the gradient difference between the ridge edge and the background.

[0024] Preferably, it also includes an Automatic Fingerprint Identification System (AFIS) system adapter output unit;

[0025] The input end of the AFIS system adapter output unit is hardware connected to the output end of the fingerprint age normalization generation unit. It is configured to convert the generated adult morphological normalized fingerprint image into the standard input format of the automatic fingerprint recognition system and output it to the automatic fingerprint recognition system to complete cross-age fingerprint comparison and candidate recall.

[0026] Preferably, the fingerprint age normalization generation unit adopts a U-Net architecture as its core skeleton, including an encoder, a decoder, and a skip connection structure;

[0027] The encoder is used to extract fingerprint orientation field and spatial deformation features, the decoder is used to reconstruct adult fingerprint images, and the skip connection structure is used to directly stitch the shallow features of the corresponding level of the encoder to the corresponding level of the decoder, providing local structural references for minutiae reconstruction and suppressing the generation of fake minutiae during the generation process.

[0028] The encoder consists of 8 convolutional layers and compresses the image resolution through 4 downsampling operations with a stride of 2; the decoder consists of 8 deconvolutional layers and reconstructs the adult fingerprint image through 4 upsampling operations with a stride of 2; the final output layer of the fingerprint age normalization generation unit normalizes the pixel values ​​to the [-1,1] range through the Tanh activation function.

[0029] (III) Beneficial Effects

[0030] Compared with existing technologies, this invention provides a children's fingerprint growth system for cross-age comparison in forensic science, which has the following beneficial effects:

[0031] 1. This pediatric fingerprint growth system for cross-age comparison in forensic science overcomes the limitation of existing explicit geometric transformation correction methods, which can only handle globally uniform deformation, by using a multi-scale encoder-decoder architecture of fingerprint age normalization generation unit and closed-loop control of multi-physical constraint loss optimization unit. The encoder extracts fingerprint orientation field and spatial deformation features, and the decoder completes the reconstruction of adult morphological fingerprint images. Multi-physical constraints are adapted to the non-uniform physiological growth law of children's fingerprints, accurately correcting local deformation and avoiding the image blurring problem caused by simple geometric magnification.

[0032] 2. This child fingerprint growth system for cross-age comparison in forensic science establishes a closed-loop hardware connection between a composite loss function constructed by a multi-physical constraint loss optimization unit and the parameter control terminal of the fingerprint age normalization generation unit. This solves the problem that general deep generation models lack constraints on the physical laws of fingerprint growth, which easily leads to the generation of false details. Through the synergistic effect of orientation field perception constraints, ridge frequency evolution constraints, and identity preservation constraints, the system locks in the invariance of fingerprint topology and suppresses the generation of fake minutiae throughout the generation process.

[0033] 3. This child fingerprint growth system for cross-age comparison in forensic science solves the problem of disconnect between the optimization target of the existing generation model and the matching logic of the automatic fingerprint recognition system by working together with the local texture discrimination unit and the multi-physical constraint loss optimization unit, and with the standardized conversion processing of the AFIS system adaptation output unit. The local texture discrimination unit performs authenticity judgment on the local ridge structure to ensure that the generated image meets the feature extraction requirements of fingerprint recognition and adapts to the high true matching rate requirements of extremely low false recognition rate in criminal investigation scenarios.

[0034] 4. This child fingerprint growth system for cross-age comparison in forensic science solves the problem that existing technologies cannot meet the stringent requirements of forensic science for the authenticity of detailed features by using a standardized preprocessing process in the fingerprint image preprocessing unit and a skip connection structure in the fingerprint age normalization generation unit. The preprocessing unit enhances the gradient difference between the fingerprint ridge edge and the background, and the skip connection provides the original local structure reference for detail reconstruction, ensuring the topological continuity of the generated fingerprint ridge and the fidelity of the microstructure. Attached Figure Description

[0035] Figure 1 This is a statistical distribution diagram of the vertical fingerprint database of the present invention;

[0036] Figure 2 This is a flowchart of the fingerprint image preprocessing unit of the present invention;

[0037] Figure 3 This is a flowchart illustrating the overall structure and closed-loop data flow of the fingerprint growth processing method for children according to the present invention.

[0038] Figure 4This is a schematic diagram of the encoder-decoder structure of the fingerprint age normalization generation unit of the present invention.

[0039] Figure 5 This is a schematic diagram of the PatchGAN architecture of the local texture discrimination unit of the present invention.

[0040] Figure 6 ROC curves for cross-age fingerprint verification using different methods of the present invention;

[0041] Figure 7 This is a comparison diagram of the feature space separation degree of different methods of the present invention;

[0042] Figure 8 This is a waterfall plot of the ablation experiment for each component of the loss function of this invention;

[0043] Figure 9 This is a graph showing the decline in matching performance of different methods of the present invention over age.

[0044] Figure 10 This is a graph showing the trend of the generated image quality index of the present invention across the age range;

[0045] Figure 11 A comparison chart of the structural domain fidelity indices of different methods of the present invention;

[0046] Figure 12 This is a diagram illustrating the white-box feature extraction and topology verification results of the fingerprint generation method of the present invention.

[0047] Figure 13 This is a comparison chart of the cross-matcher generalization performance of different methods of the present invention. Detailed Implementation

[0048] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1:

[0050] The hardware platform configuration of this embodiment is a 25 vCPU Intel Xeon Platinum 8470Q processor, 90GB of RAM, an NVIDIA GeForce RTX5090 graphics processor with 32GB of video memory, and runs on the Ubuntu 22.04 operating system, PyTorch 2.8.0 deep learning framework, and CUDA 12.8 parallel computing architecture.

[0051] The hardware connections of each unit are as follows:

[0052] 1. The output of the fingerprint image preprocessing unit is connected to the input of the fingerprint age normalization generation unit via a hardware data bus, and is used to transmit the preprocessed fingerprint image data;

[0053] 2. The first output terminal of the fingerprint age normalization generation unit is connected to the first input terminal of the local texture discrimination unit through a hardware data bus, and is used to output the generated normalized fingerprint image;

[0054] 3. The second output of the fingerprint age normalization generation unit is connected to the first input of the multi-physical constraint loss optimization unit through a hardware data bus, and is used to output the generated normalized fingerprint image;

[0055] 4. The second input of the local texture discrimination unit is connected to a real adult fingerprint standard image, and the output is connected to the second input of the multi-physical constraint loss optimization unit through a hardware data bus to output adversarial loss data;

[0056] 5. The third input of the multi-physical constraint loss optimization unit is connected to the original child fingerprint image and the standard image of real adult fingerprint. The output is connected to the parameter control terminal of the fingerprint age normalization generation unit through the hardware control bus to form a closed loop, which is used to output network parameter optimization instructions.

[0057] 6. The third output terminal of the fingerprint age normalization generation unit is connected to the input terminal of the AFIS system adapter output unit via a hardware data bus, and is used to output the final normalized fingerprint image.

[0058] The training and verification fingerprint samples used in this embodiment are from a large-scale standardized fingerprint database that is actually archived by the Chinese public security system. All samples are standard fingerprint images with a size of 640×640 pixels and a resolution of 500dpi, and have been de-identified, retaining only the fingerprint image, age at the time of collection, and time label.

[0059] The sample selection rule is to use the fingerprints of children aged 5 to 10 years as the baseline sample, and match the corresponding individual's fingerprints with longitudinal tracking samples of no more than 50 years old. The fingerprint collection time span of a single individual is controlled within 24 years.

[0060] The final valid sample contains 64,084 fingerprint images from 4,680 independent individuals, which are divided into a training set of 3,276 individuals, a validation set of 468 individuals, and a test set of 936 individuals in a ratio of 7:1:2. All fingerprint images of the same individual are grouped into the same subset, with no overlap between individuals.

[0061] refer to Figure 1You can view the complete statistical distribution of the vertical fingerprint database, including core data features such as the initial collection age distribution, the density of vertical collection times, and the age span distribution.

[0062] Example 2:

[0063] In this embodiment, the fingerprint image preprocessing unit is configured to sequentially perform standardization processing steps on the standard fingerprint image of a child received at the input terminal, and the output terminal transmits the preprocessed image to the fingerprint age normalization generation unit. The complete processing flow is referenced in [reference needed]. Figure 2 .

[0064] The specific processing steps are as follows:

[0065] Step 1. Perform histogram equalization:

[0066] Global histogram equalization is performed on the input raw child fingerprint image to expand the global grayscale dynamic range of the image, so that the faint child fingerprint ridges can be distinguished from the background in grayscale, thus solving the problem of shallow child fingerprint ridges and low contrast.

[0067] Step 2. Perform Gaussian filtering for noise reduction:

[0068] Gaussian filtering is applied to the histogram-equalized image to smooth high-frequency background noise in the image, avoid amplifying noise in subsequent sharpening operations, and protect the continuity of the fingerprint ridge structure.

[0069] Step 3. Perform Laplacian sharpening:

[0070] Laplacian sharpening is applied to the Gaussian-filtered image to enhance the gradient difference between the ridge edges and the background, improve the readability of ridge details, and output the preprocessed fingerprint image after processing.

[0071] Example 3:

[0072] This embodiment represents the core implementation of the present invention. Through the coordinated operation of the fingerprint age normalization generation unit, the local texture discrimination unit, and the multi-physical constraint loss optimization unit, it achieves the normalization conversion of a child's fingerprint to an adult morphology. The overall structure and closed-loop data flow reference... Figure 3 .

[0073] The specific implementation logic is as follows:

[0074] I. Operational Logic of the Fingerprint Age Normalization Generation Unit

[0075] The fingerprint age normalization generation unit uses a U-Net architecture as its core framework, including an encoder, decoder, and skip connection structure. For the complete structure, refer to [reference needed]. Figure 4 The unit is configured to perform the following operations:

[0076] 1. The encoder performs feature extraction: The encoder consists of 8 convolutional layers. Through 4 downsampling operations with a stride of 2, it gradually compresses the spatial resolution of the input image, extracts the orientation field flow features and spatial deformation instructions of the fingerprint in the deep feature space, suppresses the interference of high-frequency details such as dense micro ridges and sweat pores in children's fingerprints, and focuses on the core growth deformation features.

[0077] 2. The decoder performs image reconstruction: The decoder consists of 8 deconvolutional layers. Through 4 upsampling operations with a stride of 2, based on the deformation instructions output by the encoder and the closed-loop optimization parameters of the multi-physical constraint loss optimization unit, it gradually reconstructs the micro-ridge structure and texture attributes of the adult fingerprint.

[0078] 3. The skip connection structure performs feature concatenation: the output features of the four downsampling layers in the encoder are directly concatenated to the input features of the four upsampling layers in the decoder, providing a local structural reference for the reconstruction of the minutiae of the original child fingerprint. This constrains the network to learn the offset mainly in the neighborhood of the original minutiae, suppressing the risk of the model fabricating new minutiae out of thin air.

[0079] 4. Output layer performs pixel normalization operation: The final output layer of the decoder is connected to the Tanh activation function, which normalizes the pixel values ​​of the generated image to the range of [-1,1], and outputs an adult morphological normalized fingerprint image that conforms to the input standard of the automatic fingerprint recognition system.

[0080] II. Operational Logic of the Local Texture Discrimination Unit

[0081] The local texture discrimination unit adopts the PatchGAN architecture. For the complete structure, please refer to [reference needed]. Figure 5 The configuration is set to determine the authenticity of the local ridge structure between the generated normalized fingerprint image and the real adult fingerprint image, specifically by performing the following operations:

[0082] 1. Architecture configuration: The PatchGAN architecture consists of four 4×4 convolutional layers with a stride of 2 connected in series. Each convolutional layer is followed by an instance normalization layer and a LeakyReLU activation function. The input image size is 640×640 pixels. The architecture finally outputs an N×N real probability matrix.

[0083] 2. Local discrimination logic: Each element in the probability matrix corresponds to the realism probability of a 70×70 pixel local receptive field in the input fingerprint image. The continuity of the ridge structure and the authenticity of the texture in each local region are independently judged. Adversarial penalties are applied to local regions where ridge breaks, adhesions, or false textures occur.

[0084] 3. Loss Output: Based on the discrimination results between the generated normalized fingerprint image and the real adult fingerprint image, calculate and output the adversarial loss. At most physical constraint loss optimization units participate in the calculation of the total loss function.

[0085] III. Operational Logic of the Multi-Physics Constraint Loss Optimization Unit

[0086] The multi-physical constraint loss optimization unit is configured to construct a composite loss function based on the physical laws of fingerprint physiological growth, generate a total optimization parameter closed loop to adjust the network parameters of the fingerprint age normalization generation unit, and specifically perform the following operations:

[0087] 1. Calculation of Direction Field Perception Loss

[0088] The unit is configured to compute orientation field sensing loss. To lock the topological invariance of the macroscopic fingerprint pattern, its mathematical expression is:

[0089] ;

[0090] in,

[0091] ; Based on this formula, a gradient penalty is applied to pixel distortions that deviate from the original fingerprint skeleton flow direction.

[0092] 2. Calculation of Ridge Frequency Evolution Loss

[0093] The cell is configured to calculate the ridge frequency evolution loss. To ensure that the generation process conforms to the anisotropic decay law of fingerprint ridge density with age, the following steps are specifically performed:

[0094] Step 1: Using the short-time Fourier transform extraction module, obtain the local ridge frequency matrices of the generated normalized fingerprint image and the original child fingerprint image, respectively. ;

[0095] Step 2: Call the pre-stored anisotropic dynamic frequency attenuation coefficient matrix

[0096] The directional decay coefficient is the one fitted based on the spatiotemporal pattern of children's fingerprint growth.

[0097] Step 3: Calculate the frequency attenuation matrix for the current time span. Its mathematical expression is:

[0098] ;

[0099] Step 4: Calculate the frequency evolution loss based on the above matrix. Its mathematical expression is:

[0100] ;

[0101] in, These are the width and height of the fingerprint image, respectively.

[0102] The original child fingerprint image is used; the unit is based on this formula, and the constraint generation process conforms to the physical law of anisotropic growth of child fingerprints.

[0103] 3. Calculation of Loss of Identity Preservation

[0104] The cell is configured to calculate the identity retention loss. To ensure that the identity features of the generated image are consistent with the real target, its mathematical expression is:

[0105] ;

[0106] in,

[0107] This is a pre-trained fingerprint identity feature extraction network. Based on this formula, the cosine distance between the generated image and the real adult fingerprint image in the high-dimensional feature space is calculated, constraining the generation process to not destroy identity-related features.

[0108] 4. Pixel Reconstruction Loss Calculation

[0109] The unit is configured to compute pixel reconstruction loss.

[0110] Its mathematical expression is the L1 distance between the generated normalized fingerprint image and the real adult fingerprint image, which is used to constrain the pixel-level overall consistency of the generated image.

[0111] 5. Total Loss Function Calculation and Closed-Loop Optimization

[0112] The cell is configured to calculate the total loss function.

[0113] Its mathematical expression is:

[0114] ;

[0115] in, These are preset weight coefficients with different values. In this embodiment, after comprehensive optimization on the validation set, each weight coefficient is set to...

[0116] .

[0117] Based on the total loss function, the unit generates network optimization parameters through the Adam optimizer, and feeds them back to the fingerprint age normalization generation unit through a closed-loop hardware control bus to complete the iterative optimization of the network parameters. In this embodiment, the training batch size is set to 16, the initial learning rate is set to 0.0002, and the number of training iterations is set to 200.

[0118] Example 4:

[0119] In this embodiment, the AFIS system adapter output unit is configured to perform the following operations to achieve interface with mainstream automatic fingerprint recognition systems in public security criminal investigation:

[0120] 1. Format conversion: Convert the normalized fingerprint image output by the fingerprint age normalization generation unit, whose pixel value is in the range of [-1,1], into an image in the gray range of 0-255, and output it in the standard BMP format of 500dpi, which fully complies with the input standards of the public security fingerprinting and automatic fingerprint recognition system.

[0121] 2. Comparison and Adaptation: Input the converted fingerprint image into Neurotechnology VeriFinger SDK13.0. This engine strictly follows the feature extraction logic in the field of forensic science to complete fingerprint minutiae extraction, feature encoding and matching comparison, and output the matching score and candidate ranking list.

[0122] 3. Results Output: The matching and comparison results are output to the forensic science case handling process to achieve candidate recall of fingerprints across ages, narrow the scope of manual review, and provide clues to support case investigation.

[0123] Example 5:

[0124] Based on the complete configuration of Embodiments 1 to 4 above, this invention completes systematic verification on an independently isolated test set, and performs a comprehensive evaluation from four dimensions: overall matching efficiency, loss function ablation analysis, long-term robustness, and microstructure fidelity. The specific verification results are as follows:

[0125] 1. Overall matching performance evaluation

[0126] This invention is comprehensively compared with traditional geometric correction methods and general depth image generation models. The comparison methods include mainstream solutions such as Feature Alignment, TPS Warping, pix2pix, and CycleGAN.

[0127] refer to Figure 6 The ROC curves for cross-age fingerprint verification of each method can be viewed. Detailed performance data are as follows: The overall discrimination capability AUC of this invention reaches 0.999. At the extremely low false recognition rate operating point that is of great concern to forensic science, the TPR reaches 0.937 when FMR=1e-4 and 0.820 when FMR=1e-5. The feature separation degree reaches 0.557, which is significantly better than all the comparison methods.

[0128] refer to Figure 7The feature space separation of each method can be compared. After normalization, the average score of real identity matching is increased to 0.778, while the score of false matching is controlled below 0.221, achieving effective separation of the feature space.

[0129] 2. Analysis of Loss Function Ablation Experiment

[0130] To quantify the independent contribution of each component in the composite loss function, this embodiment conducted a systematic ablation test. The test was set with the constraint FMR=1e-5, and the AUC value under this condition was used as the baseline value.

[0131] refer to Figure 8 The contribution of each component to the final performance of the model can be viewed. Detailed data shows that the AUC of the baseline model relying only on pixel-level reconstruction and discriminator adversarial loss is 0.758. When the ridge frequency evolution constraint, orientation field perception constraint and identity preservation constraint are removed respectively, the model AUC all show a significant decrease. Among them, the performance loss caused by removing the identity preservation constraint is the most significant, with a performance decrease of 0.067, which verifies the core value of the multi-physics constraint mechanism.

[0132] 3. Long-term robustness assessment

[0133] refer to Figure 9 You can view the decay trend of different model matching performance as the age span increases. Traditional methods show a systematic decline in recognition rate after the age span exceeds 5 to 8 years. In contrast, the performance decay curve of this invention remains flat when the age span is extended to 24 years.

[0134] At the level of perceptual quality in image reconstruction, reference Figure 10 The changes in PSNR, SSIM, and LPIPS indices of the generated image over time can be observed. Within the test range of 0 to 24 years, the quality indicators of the generated image did not show drastic divergence and remained in a stable low range, verifying the stable adaptation capability of the invention in long-term scenarios.

[0135] 4. Microstructure fidelity assessment

[0136] In this embodiment, the VeriFinger engine is used to extract the feature point set of the generated image and the real target image. The tolerance standard for true pairing points is defined as a spatial relative Euclidean distance ≤ 1 pixel and an angular deviation ≤ 15°. Generated minutiae that do not meet this standard are judged as fake minutiae.

[0137] refer to Figure 11A comparison of the structural domain fidelity indices of various methods can be viewed. The orientation field error of the image generated by this invention is 0.08 radians, the skeleton Dice score reaches 0.89, and the minutiae F1 score reaches 0.85, all of which are significantly better than the comparison methods. In terms of controlling fake feature points, the traditional generation model produced 22.4% isolated fake minutiae, while this invention suppressed this rate to 7.8%.

[0138] refer to Figure 12 You can view the white-box feature extraction and topology verification results of the generated fingerprint. The extracted ridge skeleton is continuous and completely preserves the topology of the original pattern. There are no large areas of high-frequency fragmentation or burrs.

[0139] refer to Figure 13 You can view the generalization performance test results across matchers. The images generated by this invention achieved highly consistent matching scores on both the Neurotechnology SDK and the open-source fingerprint matching algorithm, with no significant generalization gap, and have good multi-platform adaptability.

[0140] The legal application boundary of this invention is as an auxiliary preprocessing device at the front end of the retrieval process of an automatic fingerprint recognition system, used to improve the fingerprint candidate recall capability in scenarios spanning ages from children to adults. The final identification of individual identity is completed by qualified professional fingerprint identification personnel based on the original sample and legal standard procedures. The output results of this invention are not directly used as legal evidence.

[0141] In summary, this pediatric fingerprint growth system for cross-age comparison in forensic science overcomes the limitation of existing explicit geometric transformation correction methods, which can only handle globally uniform deformation, by using a multi-scale encoder-decoder architecture of the fingerprint age normalization generation unit combined with closed-loop control of the multi-physical constraint loss optimization unit. The encoder extracts the fingerprint orientation field and spatial deformation features, while the decoder completes the reconstruction of the adult morphological fingerprint image. The multi-physical constraint adapts to the non-uniform physiological growth law of children's fingerprints, accurately corrects local deformation, and avoids the image blurring problem caused by simple geometric magnification.

[0142] Furthermore, this child fingerprint growth system for cross-age comparison in forensic science forms a closed-loop hardware connection with the parameter control terminal of the fingerprint age normalization generation unit through a composite loss function constructed by a multi-physical constraint loss optimization unit. This solves the problem that general deep generation models lack constraints on the physical laws of fingerprint growth, which easily leads to the generation of false details. Through the synergistic effect of orientation field perception constraints, ridge frequency evolution constraints, and identity preservation constraints, the system locks in the invariance of fingerprint topology and suppresses the generation of fake minutiae throughout the generation process.

[0143] Furthermore, this child fingerprint growth system for cross-age comparison in forensic science solves the problem of the disconnect between the optimization target of the existing generation model and the matching logic of the automatic fingerprint recognition system by cooperating with the local texture discrimination unit and the multi-physical constraint loss optimization unit, combined with the standardized conversion processing of the AFIS system adaptation output unit. The local texture discrimination unit performs authenticity judgment on the local ridge structure, ensuring that the generated image meets the feature extraction requirements of fingerprint recognition and adapts to the high true matching rate requirements of extremely low false recognition rate in criminal investigation scenarios.

[0144] Furthermore, this children's fingerprint growth system for cross-age comparison in forensic science addresses the problem that existing technologies cannot meet the stringent requirements of forensic science for the authenticity of detailed features through a standardized preprocessing process in the fingerprint image preprocessing unit, coupled with the skip connection structure in the fingerprint age normalization generation unit. The preprocessing unit enhances the gradient difference between the fingerprint ridge edge and the background, and the skip connections provide the original local structural reference for detail reconstruction, ensuring the topological continuity of the generated fingerprint ridge and the fidelity of the microstructure. This solves the problems in existing technologies where geometric transformation methods can only handle global uniform deformation, easily exacerbate image blurring and increase false features, and general depth generation models lack specific constraints on the physical laws of fingerprint growth, easily generate false details, and have a disconnect between the optimization target and the matching logic of the automatic fingerprint recognition system, thus failing to meet the stringent application requirements of forensic science.

[0145] The relevant modules involved in this system are all hardware system modules or functional modules that combine computer software programs or protocols with hardware in the prior art. The computer software programs or protocols involved in these functional modules are technologies known to those skilled in the art and are not improvements to this system. The improvement of this system lies in the interaction or connection between the modules, that is, in improving the overall structure of the system to solve the corresponding technical problems that this system aims to address.

[0146] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A child fingerprint growth system for forensic science cross-age comparison, characterized by, This includes a fingerprint age normalization generation unit and a multi-physical constraint loss optimization unit; The fingerprint age normalization generation unit adopts a multi-scale encoder-decoder architecture and is configured to perform inverse physical modeling of the non-uniform physical deformation of fingerprint skin with physiological age on the pre-processed standard fingerprint image of young children, and output a normalized fingerprint image corresponding to the adult form. The multi-physical constraint loss optimization unit has its input end connected to the output end of the fingerprint age normalization generation unit, and its output end forms a closed-loop hardware connection with the parameter control end of the fingerprint age normalization generation unit. The multi-physical constraint loss optimization unit is configured to construct a composite loss function based on the physical laws of fingerprint physiological growth, including orientation field perception constraints, ridge frequency evolution constraints, and identity preservation constraints. It combines pixel reconstruction loss to generate total optimization parameters, and uses closed-loop feedback to adjust the network parameters of the fingerprint age normalization generation unit, correcting the non-uniform growth deformation of children's fingerprints and suppressing forged minutiae in the generation process.

2. A forensic science age progression system for cross-age comparison of children's fingerprints as claimed in claim 1, wherein, It also includes a local texture discrimination unit; The first input terminal of the local texture discrimination unit is hardware-connected to the output terminal of the fingerprint age normalization generation unit, the second input terminal is connected to a real adult fingerprint standard image, and the output terminal is hardware-connected to the multi-physical constraint loss optimization unit. The local texture discrimination unit adopts the PatchGAN architecture, which consists of multiple convolutional layers connected in series. Each convolutional layer is followed by an instance normalization layer and a LeakyReLU activation function. The architecture finally outputs an N×N truth probability matrix. The local texture discrimination unit is configured to perform authenticity discrimination on the local ridge structure of the generated normalized fingerprint image and the real adult fingerprint image, apply adversarial penalties to ridge breaks and false textures in the local receptive field, ensure the topological continuity of the local ridges in the generated image, output the local authenticity probability matrix and adversarial loss, and transmit them to the multi-physical constraint loss optimization unit to participate in the generation of the overall optimization parameters. Each element of the probability matrix corresponds to the authenticity probability of a local receptive field of the input fingerprint image.

3. A forensic science age progression system for cross-age comparison of children's fingerprints as claimed in claim 1, wherein, The orientation field perception constraint of the multi-physics constraint loss optimization unit calculates the orientation field feature mapping of the generated normalized fingerprint image and the real adult fingerprint image respectively through the fingerprint orientation field feature extractor, calculates the mean L1 distance of the two feature mappings, applies gradient penalty to the pixel distortion that deviates from the original fingerprint skeleton flow direction, and locks the topological invariance of the fingerprint macro-pattern.

4. A child fingerprint growth system for cross-age comparison in forensic science according to claim 1, characterized in that, The ridge frequency evolution constraint of the multi-physical constraint loss optimization unit obtains the local ridge frequency matrices of the generated normalized fingerprint image and the original child fingerprint image by extracting local ridge frequency features. Combined with the anisotropic dynamic frequency attenuation coefficient matrix prefitted based on the growth law of child fingerprints, the mean L1 distance of the two frequency matrices at the corresponding spatial positions is calculated. The constraint generation process conforms to the anisotropic attenuation law of ridge density of child fingerprints with age.

5. A child fingerprint growth system for cross-age comparison in forensic science according to claim 1, characterized in that, The identity preservation constraint of the multi-physics constraint loss optimization unit extracts high-dimensional identity features from the generated normalized fingerprint image and the real adult fingerprint image through a pre-trained fingerprint identity feature extraction network, calculates the cosine distance between the two high-dimensional identity features, and constrains the identity-related features of the generated image to be consistent with the real target image.

6. A child fingerprint growth system for cross-age comparison in forensic science according to claim 2, characterized in that, The total loss function of the multi-physics constraint loss optimization unit is obtained by weighting and summing adversarial loss, pixel reconstruction loss, orientation field perception loss, identity preservation loss, and ridge frequency evolution loss through preset weight coefficients with different values.

7. A child fingerprint growth system for cross-age comparison in forensic science according to claim 1, characterized in that, It also includes a fingerprint image preprocessing unit; The output of the fingerprint image preprocessing unit is hardware-connected to the input of the fingerprint age normalization generation unit. It is configured to sequentially perform histogram equalization, Gaussian filtering, and Laplacian sharpening on the input standard fingerprint image of a child, and output the preprocessed fingerprint image. The histogram equalization is used to expand the global grayscale dynamic range of the fingerprint image, the Gaussian filtering is used to smooth the high-frequency noise of the image, and the Laplacian sharpening is used to enhance the gradient difference between the ridge edge and the background.

8. A child fingerprint growth system for cross-age comparison in forensic science according to claim 1, characterized in that, It also includes an Automatic Fingerprint Identification System (AFIS) system adapter output unit; The input end of the AFIS system adapter output unit is hardware connected to the output end of the fingerprint age normalization generation unit. It is configured to convert the generated adult morphological normalized fingerprint image into the standard input format of the automatic fingerprint recognition system and output it to the automatic fingerprint recognition system to complete cross-age fingerprint comparison and candidate recall.

9. A child fingerprint growth system for cross-age comparison in forensic science according to claim 1, characterized in that, The fingerprint age normalization generation unit adopts the U-Net architecture as its core skeleton, including an encoder, a decoder, and a jump connection structure; The encoder is used to extract fingerprint orientation field and spatial deformation features, the decoder is used to reconstruct adult fingerprint images, and the skip connection structure is used to directly stitch the shallow features of the corresponding level of the encoder to the corresponding level of the decoder, providing local structural references for minutiae reconstruction and suppressing the generation of fake minutiae during the generation process. The encoder consists of 8 convolutional layers and compresses the image resolution through 4 downsampling operations with a stride of 2. The decoder consists of 8 deconvolutional layers and reconstructs an adult fingerprint image through 4 upsampling operations with a stride of 2. The final output layer of the fingerprint age normalization generation unit normalizes the pixel values ​​to the [-1, 1] interval through the Tanh activation function.

Images